Enhancing Echocardiographic Segmentation through KL-Divergence-Based Intensity Distribution Constraints in U-Net Models

Incorporating prior knowledge such as constraints can be crucial to improving the performance of image analysis methods, particularly when dealing with corrupted images, poor quality images, low contrast, and a lack of training data. To our knowledge, there has been nothing in the literature using the KL-divergence between the density distributions of the segmented objects as a constraint. Instead, the KL-divergence is used as a loss function between the predicted image and its label. In the present paper, we try to demonstrate the efficiency of applying an intensity distribution of the region of interest as a constraint with the KL-divergence function in 2D echocardiographic imaging segmentation (left ventricle segmentation). For this, we use the U-net neural network to which we added a second pseudo output which serves as a constraint, trained on echocardiographic medical images of good, medium and poor quality. A two-point improvement was achieved by applying our constraint, when compared with the use of the cross-entropy alone. For images where the cross-entropy performs better than our method, imposing a constraint provides a smoother and more logical segmentation, with a shape that more closely resembles the label than that obtained by only using the cross-entropy. In the segmentation of medical ultrasound images, the use of an intensity distribution as a constraint can be highly beneficial, especially when the target region has an intensity distribution different from that of the background distribution.